MindBase synthesis

Type: agentic-system-analyses/types/agentic-system-synthesis.md

Bounded synthesis

From captured sources to retained knowledge

MindBase is a model-maintained knowledge system whose persistent products include project context, research pages and source files. Later builds, questions and hosted agent sessions can consume those products. This account concerns https://github.com/frankchu91/mindbase-llm-wiki at c5273d2336c4d686df41c2a1495f14f085707f51, with an applicability cutoff of 2026-11-08. The evidence is source code and shipped instructions, not an executed installation, user corpus or experiment. The boundary is complete artifact, partial loop: the repository is available, but its material operational paths are not exhaustively traced. External model-provider internals and host scheduling, permissions and compaction are not MindBase implementation evidence (boundary, Source register).

The system has several execution owners, not one universal agent loop. The Express server coordinates fixed operations and background work; a configured model proposes content or requests file reads. A separate MCP server exposes persistence and retrieval tools. Plugin commands ask an external agent host to dispatch restricted workers. Markdown carries interpreted content; action envelopes, paths, indexes, hashes and verdicts have fixed-rule consumers. Search uses keyword and graph access state, with a separate retained dense-vector cache. These access structures influence selection without authenticating the selected content (RT-OBJ-actions, MEM-OBJ-access-metadata, MEM-OBJ-embeddings).

A web contribution illustrates the ordinary progression. The server first captures the supplied text, unless an existing source path is supplied. It then gives a model project guidance, current context and budgeted related-page hints. Generated JSON undergoes structural validation and at most one repair completion. The server filters duplicate slugs and stores a ten-minute in-memory plan. The human-facing UI exposes individual action checkboxes; Apply sends selected indexes. The server consumes the stored plan and executes those actions sequentially. Source capture therefore precedes wiki approval, and later action failure does not undo earlier writes. The apply branch does not verify current-project or requesting-user ownership of the stored plan, so this is a selection protocol, not an authorization guarantee (RT-RTE-contribute-plan).

Admission differs on other routes. Server build and research apply structurally valid generated actions without that checkbox stage. Configured question answering can save marked output automatically. Direct tree editing accepts caller-supplied bodies. CLI ingest instead offers batch confirmation, with a --yes alternative. These are concrete limits on the README's claim that every AI change is human-approved, not evidence that users never approve through an external host (RT-CLM-approved-changes, RT-RTE-build-context, RT-RTE-research-page, RT-RTE-ask-answer, MEM-RTE-tree-edit, MEM-RTE-cli-ingest).

Building context and delivering the result

Server build selects at most thirty contributor/research files newer than context.md, taking up to 4,000 characters per body. It does not select raw extraction files through this gather function. A newer context timestamp is not a per-source incorporation receipt: capped-out older files can fall behind the next cutoff. Full README, existing context and scaffolding are additional inputs, so the excerpt limits are not a total invocation budget (RT-RTE-build-context, memory build annotation).

Context controls are branch-specific. Server replacement rejects over 400 lines and snapshots prior nonempty content; section append checks the resulting line count without snapshotting. But the common action schema also admits add_wikilinks targeting context.md, whose missing-link append has neither check nor snapshot. That is a reachable implementation alternative within the ops executor, not an observed overflow or model choice. Direct tree editing is an additional unchecked writer. The MCP replacement tool separately relocates overflow into a research page. Neither the executor nor all writers collectively enforce a universal context cap (RT-RTE-build-context, MEM-OBJ-context, MEM-RTE-tree-edit).

The representative multi-step hosted workflow is /mb:build. The foreground agent is instructed to resolve a project and delegate its ID to a builder. The builder should validate structure, load project files, gather newer source paths, Read their bodies, synthesize replacement context, write through MCP, rebuild the index and append a log. Only a one-line summary is requested back. This keeps detailed source work in a child context by prescription, while persistent context/research writes escape through disk for later consumers. Actual host isolation and tool-grant enforcement remain unobserved (RT-RTE-plugin-build, workflow account).

The builder's structural check reports missing paths; it is not a semantic synthesis test or a token required by the writer. Instructions to preserve prior knowledge and flag contradictions are not checked against the replacement body. The MCP tool uses temporary-file rename for context replacement, but overflow writing, logging and index refresh are separate effects. Snapshot errors can be swallowed, and no automatic restore or end-to-end transaction is established. Delivery relies on the child's summary rather than an independent receipt certifying every step. The workflow is afforded by shipped instructions; individual handlers are implemented; successful execution and semantic preservation are not demonstrated (workflow account, workflow annotation).

Reuse, diagnosis and the scope of reliance

Core question answering retrieves retained material for a question. By default it selects six ranked candidates, initially supplies at most three bodies clipped to 3,000 characters within a source budget, and includes the last ten history messages. A model can request further files through a bounded tool loop. The source budget does not cover all messages, full hot context/index or subsequent file results; removed history is not summarized by this function. Separately, the session hook pushes the selected project's whole README/context/index on startup, clear or compact, without question-specific filtering or truncation. This reloads persisted project knowledge, not arbitrary evicted conversation turns (RT-RTE-ask-answer, RT-BAP-session-context, memory read-back).

These consumers do not all share the same retrieval implementation. MCP answering expands keyword seeds through the graph and reconstructs legacy note paths; it is not core askQuestion. Chats and separately filed excerpts have independent retention and withdrawal. Deleting a transcript does not establish withdrawal of its excerpt or every derived mention (MEM-RTE-mcp-retrieval, MEM-RTE-chat-retention).

Checks support narrower conclusions than truth labels suggest. Action schemas check executable shape. Topic synthesis checks citation coordinates and conditionally prunes sentence fragments; citation-free prose can survive when the thread has valid citation metadata. Ops lint requires normalized quote occurrences for contradiction/stale findings, but two verified entries can repeat the same page and quote. These checks can reject malformed references or unsupported quotation occurrences; neither proves that a sentence follows from its source or that two quotations conflict (EPI-RTE-topic-synthesis, EPI-RTE-evidence-quotes).

A separate contradiction probe judges selected page pairs, caches a three-way verdict and reason, and turns cached contradiction enums into ranked review suggestions. Error fallbacks can also yield unrelated; cache reuse is not fresh semantic evaluation. Ops lint and topic reports do not themselves repair criticized knowledge. A different health API can crosslink and dispatch model-generated L2 rewrites directly, with handler-local vetoes for human-touched content. The curator's requested/granted health-tool name differs from the registered API name, so curator-to-writer reachability is unestablished. This is neither proof of a working curator repair path nor a host-safety guarantee (EPI-RTE-contradiction-probe, EPI-RTE-probe-consumption, EPI-RTE-health-maintenance).

What the retained changes establish

The strongest supported trace-fed contribution is conditional automatic answer filing: supplied conversational history enters generation; with autoSave enabled and matching output, software retains answer text that later questions can retrieve. This is a wired route from session experience to durable later context, not merely raw transcript storage. It does not require a parameter update or novel proposition to qualify. No supplied run shows that branch operated, and no comparison establishes improved future capacity. Other source-to-context updates have mixed or unknown original provenance, so this positive does not classify all synthesis as trace-derived (RT-RTE-ask-answer, profile verification).

Project README guidance also reaches future proposal models as instructions. Persistent content revision and later guidance are real capacities, but the accepted account does not establish content-directed criticism and iterative revision of the system's method. It establishes no produced accepted ampliative claim, warranted formal derivation or post-acceptance integration instance in the inspected routes; actual model transformations remain semantically indeterminate without source/output pairs. The runtime provides no separate complete assessment of theory-builder conditions 1–4, and these findings do not license an autonomous theory-builder verdict. Computational proposal and progression roles are supported independently of correctness. Reflection, improved capacity through criticism, and exercised self-improvement are separately unestablished—not refuted by missing run evidence or by fixed provider weights (RT-BAP-project-rules, epistemic bounded conclusion, report verification).

For retaining material and making it available to later questions or sessions, the records establish concrete persistence and delivery mechanisms. For exhaustive source incorporation, consistently warranted knowledge or the claim that context gets sharper with every contribution, they do not establish the necessary coverage or benefit. Source/output pairs with provenance and evidence-consuming acceptance decisions would change the semantic assessment. Traces tying retained revisions to later decisions, with an appropriate comparison, would change the capacity or self-improvement assessment. Uniform controls across alternate writers, per-source incorporation receipts and complete invalidation wiring would change the corresponding implementation guarantees (EPI-CLM-knowledge-compounds, memory integration issues).

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Requested filing and editable schema instructions lack declared route coverage.	SRC-1; RT-RTE-ask-answer, MEM-RTE-chat-retention, MEM-RTE-tree-edit, report-verification limits	Verifier inspected answer-file-back, its plan builder and schema replacement/reset; analyst inventory remains partial.	Complete filing, admission or instruction-edit comparisons. Verifier inspection does not create canonical analyst routes.	Accepted records covering default approvals, note/edge effects and schema replacement/reset with their actual consumers.
Retained classification guidance and examples lack declared memory/epistemic inventory coverage.	SRC-1; memory coverage, epistemic coverage, report-verification limits	Verifier inspected classification metadata writes, retained rules, up to ten recent titles per folder and clipped note inputs.	Complete working-context, retained-authority and policy-adaptation inventories; extrapolation from build/hook consumers to classification.	Canonical records for guidance/examples, selectors, admission and later classification effects.
Other included workers, callers and executor alternatives remain partly uninspected.	SRC-1; MEM-RTE-cli-ingest, EPI-RTE-health-maintenance, memory coverage	Complete repository artifact, partially traced capture/OCR/RSS, brief/SRS, compiler, client and evaluator paths.	Exhaustive lifecycle, trace-fed update or evaluation inventories; system-complete negatives.	Source-grounded records tracing material producers, admission and downstream consumers on those alternatives.
No observed operation, user corpus or controlled comparison is registered.	SRC-1; RT-RTE-build-context, RT-RTE-ask-answer, EPI-CLM-knowledge-compounds	Code and doctrine/design at the frozen commit only.	Successful synthesis, original source warrant, exercised history-fed filing, improved future capacity or causal component benefit.	Provenanced source/output pairs and execution traces; appropriately controlled comparisons for improvement or component attribution.
Hosted order, isolation, permissions and recovery depend on external behavior.	SRC-1; RT-RTE-plugin-build, workflow account, EPI-RTE-health-maintenance	Shipped command/agent texts and MCP handlers; curator/API name mismatch unresolved.	Enforced build sequence, actual context isolation, guaranteed backups, end-to-end completion or curator writer reachability; also prevents inferring host safety from the mismatch.	Host/tool-name mapping and execution evidence showing grants, sequencing, snapshot outcomes, delivery and recovery.
Layout compatibility, invalidation and cascading withdrawal remain unresolved across writers.	SRC-1; MEM-RTE-mcp-retrieval, MEM-OBJ-access-metadata, MEM-OBJ-embeddings, MEM-RTE-synthesis-cache, MEM-RTE-tree-edit	Legacy-path reads, source hashes, stale queues, reindexing and independent file deletion/restoration routes.	Uniform v2 recall, fresh cached answers or complete removal of derived material after edits/deletion.	Complete mutation-to-cache dependency records, layout-aware consumers and evidence of invalidation/withdrawal propagation.
Timestamp selection and branch-local controls do not certify incorporation or a global size limit.	SRC-1; RT-RTE-build-context, MEM-OBJ-context, MEM-RTE-tree-edit	Capped server source gathering, hosted mtime gather, checked replacements/appends, unchecked wikilink append and tree writes.	Every source incorporated, executor-wide 400-line cap or universal snapshot protection.	Per-source receipts and controls covering every relevant writer; operation evidence for actual incorporation.

Source-read log: none. This synthesis uses the supplied accepted members only.